Adobe bets on AI agents to automate creative work

The company pitches “agentic” assistants as time-savers, while a staggered rollout hints at hedging.

Adobe says AI will handle the dull parts of creativity; the product calendar tells a more cautious story. At its MAX conference in Los Angeles, Adobe unveiled AI assistants for Photoshop and Adobe Express that execute multi-step edits from a chat box, batch routine fixes, and offer personalized recommendations—“agentic” systems built to complete tasks, not just generate assets.

The Breakdown

• Photoshop's AI assistant (private beta) chains multi-step edits from prompts; Express version launches in public beta with uneven timing signaling caution.

• Firefly Image Model 5 adds native 4MP generation and layered editing; third-party models from Google and Black Forest integrate alongside Adobe's proprietary systems.

• Custom model training lets creators build character and style consistency using 6-12 images; rollout scheduled for year-end with closed beta starting now.

• Generate Soundtrack creates licensed instrumentals while Generate Speech adds voiceovers; Project Moonlight and ChatGPT integration remain early-stage teasers without firm timelines.

What’s actually new

Earlier tools asked users to know the knobs to turn. Generative Fill could remove a pole; Firefly could conjure a skyline. But you still had to navigate menus and manage layers.

The assistants flip that flow. In Photoshop’s agentic mode, the interface shrinks to a prompt: “Remove the background and increase saturation.” The system executes both steps, then lets you jump back to sliders and masks for fine control. Express works similarly, turning conversational instructions into chained edits. This is different.

Under the hood, Firefly Image Model 5 upgrades native output to 4-megapixel resolution and adds layered, prompt-based editing. Upload an image and the model identifies distinct elements—a fence, chopsticks—so you can move, resize, or replace them while it adjusts shadows and lighting to keep the composition coherent. It’s an explicit push from “paint what you want” to “edit what exists.”

Adobe is also opening doors. Photoshop’s Generative Fill now supports third-party models—Google’s Gemini 2.5 Flash and Black Forest Labs’ FLUX.1 Kontext—alongside Firefly. Users can cycle through results and pick what looks best. Topaz technology powers a new Generative Upscale for pushing low-res images to 4K. Premiere gets AI Object Mask to auto-isolate people and objects for targeted color grading without manual rotoscoping. Lightroom’s Assisted Culling ranks huge shoots by focus, angle, and sharpness to nominate keepers. Fewer clicks, more throughput.

Audio joins the party. Generate Soundtrack creates licensed, royalty-safe instrumentals that sync to video, while Generate Speech—leveraging Firefly and ElevenLabs—produces voice-overs in 15 languages with adjustable emotional tags. It’s meant to be safe to use on platforms that police copyright. That matters.

The custom-model contradiction

Adobe is pitching empowerment and control in the same breath. Custom models let individuals and teams train Firefly on their own art—roughly 6–12 images for a character, a bit more for a tone—so they can keep visual identity consistent across projects without starting from scratch. Commercial safety is the sales hook.

But training your style inside Firefly also deepens platform lock-in. Move to another toolchain and you’ll have to rebuild that stylization. At the same time, letting Google and Black Forest models run inside Photoshop concedes that “best result wins,” not “house model everywhere.” It’s a platform play. And a hedge.

The strategic read: Adobe wants to own the workflow layer while staying model-agnostic when it helps quality. No one provider dominates image generation the way Google dominates search. Owning the switchboard is safer than betting the studio on one in-house model. Smart. Also revealing.

Who Adobe is targeting

Adobe’s VP of generative AI describes the audience as “next-generation creative professionals,” people comfortable with GenAI who don’t want to memorize decades of Photoshop lore. That’s a different buyer than the shortcut ninja who can dodge and burn in their sleep.

The risk is obvious. Adobe’s moat has long been muscle memory and file fidelity inside big-budget workflows. Over-simplify the front end and veterans may worry about losing deterministic control. The assistants try to bridge that gap: power users can drop to layers and curves; newcomers can ask for “punchier color, softer skin, cooler shadows.” Accept the trade-off. “Good enough” automation will be fine for social and marketing content. For cinema color or high-end retouching, trust must be earned.

The rollout signals

The timeline says caution. Express gets a public beta because stakes are lower. Photoshop’s assistant stays private, because a bad automated action in a pro pipeline costs money. Firefly Image Model 5 is “in the months to come”; some coverage pegs widespread availability closer to 2026, while partner models work in Photoshop now. That sequencing keeps quality and safety claims intact.

There’s more in the teaser tier. Adobe previewed Express running inside ChatGPT and a “Project Moonlight” concept to carry style and context across apps and connected social accounts. Both are early. Both could be sticky if they work.

Just as notable is what Adobe didn’t say. No pricing for custom models. No clarity on compute costs tied to higher-res generation and layered edits. No explicit licensing parity for outputs from third-party models inside Adobe apps versus Firefly’s “commercially safe” promise. Those details decide budgets.

What this means for rivals

The logic extends beyond Adobe. If assistants reliably run multi-step workflows, the advantage shifts from knobs to outcomes—and from model supremacy to orchestration UX. Canva, Figma, CapCut, and the rest will need agents that understand sequences, not just prompts. Ecosystem openness helps, even when it stings the house model’s pride.

For enterprises, the appeal is operational: faster turnarounds, consistent style across teams, fewer hours lost to tedious edits. The risk is governance. Who checks that the automated sequence didn’t introduce artifacts, bias, or licensing landmines? Process, not magic, will make or break adoption. Ship carefully.

What to watch next

Three questions will sort hype from habit. First, does the Photoshop assistant reliably do what pros intend, on deadline, with version control that plays nicely in teams? Second, do custom models actually reduce revision cycles for brand work, or do they spawn new QA overhead? Third, does ecosystem openness accelerate quality, or does “pick a model” become another decision tax in fast workflows? Answers will show up in invoices. Quickly.

Why this matters

Automation is moving from prompts to procedures. If agents can chain edits as well as humans, the creative stack re-organizes around outcome-driven workflows, not tool mastery.
Control is shifting to the platform layer. Adobe’s embrace of third-party models suggests that owning orchestration and safety beats owning every model—and competitors will follow.

❓ Frequently Asked Questions

Q: What does "agentic AI" mean and how is it different from regular AI tools?

A: Agentic AI completes multi-step tasks autonomously rather than just responding to single prompts. Instead of removing a background then separately adjusting saturation, you tell the assistant "remove the background and increase saturation" and it executes both steps in sequence. Traditional AI tools generate content; agentic systems execute workflows.

Q: When will Photoshop's AI assistant actually be available to use?

A: Photoshop's assistant is currently in private beta with waitlist access only. Adobe Express's version launched in public beta immediately. Firefly Image Model 5 arrives "in the months to come"—likely 2026 based on the April 2025 launch of Model 4. Custom model training enters closed beta now, with broad release by year-end 2025.

Q: How much will custom model training cost?

A: Adobe hasn't announced pricing yet. The feature enters closed beta for individual creators and businesses at the end of 2025, but no subscription tier details or per-model costs have been shared. This matters for budget planning, particularly for agencies managing multiple client brands that would need separate trained models.

Q: Are images created with Google and Black Forest Labs models commercially safe like Firefly outputs?

A: Adobe hasn't confirmed whether partner model outputs carry the same commercial licensing guarantees as Firefly-generated content. This uncertainty matters for professional use—Firefly's "commercially safe" promise has been a key selling point. Without explicit licensing parity, creators may face copyright risks using third-party models through Adobe's interface.

Q: How does custom model training actually work?

A: You upload 6-12 images to train Firefly on a specific character, or slightly more for a visual tone or style. The system uses Adobe's base Firefly model as the foundation, trained on licensed data, then fine-tunes it on your examples. This keeps outputs commercially safe while maintaining consistent characters or brand aesthetics across projects.